Category:

Neural Network

Generative AI in the Automotive Industry

by Massimo 2024-05-09

As the global automotive industry navigates a transformative era marked by rapid technological advancements and shifting market dynamics, leading companies like Toyota are strategically pivoting to remain competitive. The Japanese automaker has forecast a 20% decline in annual profits as it increases investments in electric vehicles (EVs) and artificial intelligence (AI) to counter intensifying competition, particularly from Chinese firms. This shift highlights a significant trend: the growing adoption of generative AI technologies in automobile manufacturing and supply chain management.

Generative AI is revolutionizing the automotive sector with its ability to create, simulate, and iterate models rapidly. This technology is proving invaluable in designing vehicle components, optimizing assembly lines, and enhancing the overall manufacturing process. AI-driven algorithms, for example, can quickly generate multiple iterations of a vehicle’s part design, identifying the most efficient and cost-effective options. This capability significantly reduces the limitations typically associated with traditional prototyping due to physical and financial constraints.

AI’s potential extends to mitigating manufacturing bottlenecks before they lead to disruptions. By analyzing data from various stages of the manufacturing process, AI models can forecast potential delays and suggest proactive adjustments to machinery or workforce deployment, ensuring smoother operations.

The integration of AI into established industries, however, is accompanied by challenges. Data privacy, cybersecurity, and substantial initial investments pose significant barriers. Additionally, there is a crucial human aspect—ensuring that the workforce is adequately trained to work alongside sophisticated AI systems is essential for seamless integration.

Despite these challenges, the benefits—such as higher efficiency, reduced costs, and improved product quality—make the pursuit of AI technologies worthwhile. As demonstrated by Toyota’s strategic shift, significant investments in AI and EVs are essential for future-proofing operations against an increasingly competitive backdrop.

AI use cases for automobile industry

Design Optimization and Customization

Generative AI can revolutionize the design process in the automotive industry by enabling rapid prototyping and testing, allowing manufacturers to explore a broader range of design options before finalizing the most functional and cost-effective models. AI also facilitates mass customization, allowing customers to personalize features such as colour, interior materials, and component designs, which enhances customer satisfaction and engagement.

Traditional vehicle design processes are indeed time-consuming and costly, often involving extensive physical prototyping and testing. Moreover, consumer demand for personalized and unique vehicle features has risen, putting pressure on automakers to offer a broader range of options without significantly increasing production costs or time.

Deploy generative AI technologies to automate and enhance the design process. This involves using AI algorithms to generate multiple design options based on predefined parameters and constraints, quickly iterating through possibilities to identify the most effective designs. AI can also be employed to enable mass customization options for consumers, allowing them to select various features, such as exterior color, interior materials, and additional accessories, which are then automatically integrated into the vehicle’s final design.

Example of implementation of a generative AI solution for car design customization

Data Collection: Gather extensive data on current design parameters, consumer preferences, and manufacturing capabilities.
AI Model Development: Develop or adapt AI algorithms capable of generating vehicle designs and accommodating customization requests.
Integration with CAD Software:
Integrate AI with existing computer-aided design (CAD) systems to automate the translation of AI-generated designs into actionable blueprints.
Use rapid prototyping tools to create physical or digital prototypes of selected designs for testing and validation.
Create an intuitive consumer interface that allows buyers to customize their vehicles online, which directly feeds into the AI system for real-time design adaptation.
Implement a feedback system to continuously improve the AI design models based on consumer satisfaction and manufacturing feedback.

Supply Chain Resilience

Generative AI can significantly enhance supply chain management by predicting potential disruptions and suggesting optimal responses. This includes managing inventory levels dynamically, optimizing route planning for logistics based on real-time traffic data, weather conditions, and other variables, and identifying risk mitigation strategies by simulating various scenarios. By improving forecasting accuracy, AI can help reduce overstock and shortages, thereby increasing operational efficiency and reducing costs.

Predictive Maintenance

Using generative AI, automakers can implement predictive maintenance for vehicles and manufacturing equipment. AI algorithms analyze data from onboard sensors to predict when a vehicle or a part is likely to fail, enabling pre-emptive maintenance that can prevent costly repairs and downtime. This not only extends the lifespan of the vehicles but also enhances customer trust and satisfaction by improving reliability and safety.

Enhanced Quality Control

AI-driven systems can continuously learn from various stages of the manufacturing process and can be trained to identify defects or anomalies in materials or assembly. This application of generative AI can significantly enhance quality control processes, ensuring that every product meets stringent quality standards. The ability to detect and rectify issues early in the manufacturing process can lead to significant savings and maintain brand reputation for high quality.

Autonomous Vehicle Development

Generative AI can play a pivotal role in the development of autonomous vehicles. By using AI to simulate countless driving scenarios, manufacturers can train and test autonomous vehicle systems under a wide range of conditions, including extreme ones that are rare or dangerous to replicate in real life. This extensive testing helps improve the safety and reliability of autonomous vehicles, accelerating their development and readiness for consumer markets.

2024-05-09

Artificial Intelligence Data Science and Governance Neural Network R&D

Find cancer with AI: a closer look at CT scan analysis with Self-Supervised Learning (SSL)

by Massimo 2024-04-08

Self-Supervised Learning (SSL) is a fascinating area of machine learning, especially if you’re into the tech world or curious about how artificial intelligence (AI) learns to make sense of the world around it. Imagine being dropped in a new city with no map, no guide, and the task to find your way to a famous landmark. You’d start noticing patterns, like how certain signs point towards central areas, or how the flow of people might lead you to popular spots. This is a bit like how Self-Supervised Learning (SSL) works. In the tech world, we often have a ton of data but not enough labels or explanations for what that data means.
Labelling data can be like giving a tourist a detailed map – very helpful, but expensive and time-consuming to create. SSL skips the map: it teaches AI models to learn from the data itself, finding its own patterns, and creating its own understanding of what things mean. It’s a bit like teaching yourself Italian by watching movies without subtitles; you pick up on cues, context, and repeat expressions until things start to make sense.

Reflecting on my early days in the R&D industry, I recall how SSL seemed like a distant concept, almost esoteric in its nature. But diving deep into its mechanisms, I began to see its practicality unfold in real-time, transforming abstract data into meaningful patterns. From my own experience diving into the depths of machine learning, I’ve found SSL to be akin to the thrill of solving a complex puzzle without having all the pieces from the start. It’s a journey of discovery that continuously fascinates me. I still remember the ‘aha’ moment when I first truly understood the power of SSL. I was working late one evening, experimenting with an SSL algorithm on a seemingly inscrutable dataset. Suddenly, patterns began to emerge, revealing insights that had been hidden in plain sight. It was a vivid reminder of why I fell in love with the field of AI in the first place.

Having spent countless hours tweaking algorithms and adjusting models, the shift towards applying SSL in healthcare, particularly in image detection, felt like stepping into a new world of possibilities. It was a convergence of my passion for machine learning and a deeper mission to contribute to societal well-being.

The battle against cancer leverages Computed Tomography (CT) scans for early and accurate diagnosis. Despite their utility, interpreting these scans remains a highly skilled task. Variabilities in tumour appearances and the subtlety of early signs present significant challenges. Self-Supervised Learning emerges as a promising solution. It thrives on unlabelled data, learning from the input to uncover intrinsic patterns. SSL’s application in cancer identification via CT scans capitalizes on its ability to process vast volumes of unlabelled data efficiently. This approach offers cost and time savings, improves model generalization, and enhances feature extraction. However, SSL’s computational demands, potential for overfitting, and challenges in model interpretability warrant attention.

Integration of SSL in CT scan analysis

The application of SSL in CT scan analysis for cancer identification relies on its ability to harness large volumes of unlabelled data. Each SSL application offers a unique advantage in the identification of cancers through CT scans, leveraging the vast amounts of unlabelled data available in medical imaging. Combining these techniques together can improve the possibility to recognize cancers at the best. Let’s see some application of SSL:

Reconstruction Error Modelling for Anomaly Detection: Implementing SSL through reconstruction error modelling involves teaching a model to reconstruct CT scan images. The model learns to identify normal anatomical structures by minimizing the difference between the original and reconstructed images. Anomalies such as tumours manifest as significant reconstruction errors since the model is less familiar with these features, highlighting potential areas of concern. This technique leverages unlabelled CT scans to train models in distinguishing between normal and abnormal tissues, enhancing early cancer detection capabilities.
At the end of this post I put an implementation example of this detection using Transformers.
Predictive Coding for Feature Representation: One particular project that stands out in my memory involved the use of predictive coding for feature representation. The complexity of the data was overwhelming at first, but as the model began to unveil the subtle nuances between healthy and cancerous tissues, the potential of SSL in making a tangible difference in people’s lives became incredibly clear. Utilizing predictive coding frameworks in SSL enables models to predict parts of a CT scan based on surrounding information. By learning to anticipate the appearance of adjacent tissues, the model develops a deep understanding of normal anatomical patterns. Deviations from these predictions can indicate the presence of cancerous lesions. This method uses vast quantities of unlabelled data to refine its predictive accuracy, improving the model’s ability to detect subtle signs of cancer that may be overlooked in standard reviews.
Contrastive Learning for Tumour Characterization: Applying contrastive learning involves training a model to identify similarities and differences between pairs of CT scan images. By comparing scans with known cancerous lesions to those without, the model learns to distinguish between cancerous and non-cancerous tissues. This approach is particularly valuable in enhancing the model’s ability to recognize various types of cancer, as it leverages unlabelled data to learn from a broad spectrum of cases, thereby improving diagnostic precision.
Temporal Sequence Analysis for Cancer Progression Monitoring: Leveraging SSL for temporal sequence analysis involves analysing sequential CT scans of a patient over time. By learning the typical progression of anatomical changes, the model can identify deviations that suggest the development or progression of cancer. This technique allows for the monitoring of high-risk patients by utilizing existing CT scan data, offering a proactive approach to cancer care without the need for explicit labelling.
Generative Modelling for Synthetic Lesion Generation: Employing generative adversarial networks (GANs) within an SSL framework to create synthetic images of CT scans with various types of lesions. These synthetic images can be used to augment existing datasets, particularly when certain types of cancer are underrepresented. By training models on a more diverse set of images, including those generated to mimic rare cancers, SSL can significantly improve the robustness and generalizability of cancer detection algorithms, ensuring that they are well-equipped to identify a wide range of cancer types across different patient demographics.

Drawing from my years of working with both supervised and self-supervised learning models, I’ve observed that while SSL can drastically reduce the need for labelled data, it requires a nuanced understanding of data patterns to truly excel.
It’s a bit like learning to read between the lines of a complex novel, gaining insights that are not immediately obvious.

Transformers in "Reconstruction Error Modelling for Anomaly Detection"

The backbone of our approach is a convolutional autoencoder (sequence to sequence transformer), a type of neural network that learns to encode input data into a compact representation and then decode it back to the original form.

The architecture is composed of two main parts:

Encoder: a Convolutional Neural Network. This component compresses the input CT scan into a lower-dimensional latent space. It consists of several convolutional layers followed by pooling layers that progressively reduce the spatial dimensions. (see the Computer vision concept of CNN).
Decoder: This part attempts to reconstruct the input image from the latent representation. It mirrors the encoder architecture but uses transposed convolutional layers to upsample the features back to the original image dimensions.

The training process involves minimizing the reconstruction error, which is the difference between the original CT scan and its reconstructed version produced by the decoder. The objective is to make the model sensitive to anomalies.

Anomaly Detection

After training, the model’s performance in anomaly detection hinges on its response to new, unseen CT scans:

Reconstruction Error as Anomaly Indicator: For a new CT scan, the model will attempt a reconstruction. If the scan contains anomalous features, like a tumour, the model, which is trained primarily on normal anatomy, will likely reconstruct these areas poorly. The reconstruction error in these regions will be significantly higher than in normal areas.
Thresholding: To automate the detection, we establish a threshold for reconstruction error. Areas in a scan where the error surpasses this threshold are flagged as potential sites of anomalies.

A quick implementation

import torch
import torch.nn as nn
import torch.optim as optim
from torch.utils.data import DataLoader, TensorDataset
import torchvision.transforms as transforms
from torchvision import datasets

class ConvAutoencoder(nn.Module):
    def __init__(self):
        super(ConvAutoencoder, self).__init__()
        
        # Encoder layers
        self.encoder = nn.Sequential(
            nn.Conv2d(1, 16, 3, stride=1, padding=1),  # Output: (16, 256, 256)
            nn.ReLU(),
            nn.MaxPool2d(2, 2),  # Output: (16, 128, 128)
            nn.Conv2d(16, 8, 3, stride=1, padding=1),  # Output: (8, 128, 128)
            nn.ReLU(),
            nn.MaxPool2d(2, 2),  # Output: (8, 64, 64)
            nn.Conv2d(8, 8, 3, stride=1, padding=1),  # Output: (8, 64, 64)
            nn.ReLU(),
            nn.MaxPool2d(2, 2)  # Output: (8, 32, 32)
        )
        
        # Decoder layers
        self.decoder = nn.Sequential(
            nn.ConvTranspose2d(8, 8, 3, stride=2),  # Output: (8, 65, 65)
            nn.ReLU(),
            nn.ConvTranspose2d(8, 16, 3, stride=2),  # Output: (16, 129, 129)
            nn.ReLU(),
            nn.ConvTranspose2d(16, 1, 3, stride=2, output_padding=1),  # Output: (1, 257, 257)
            nn.Sigmoid()
        )

    def forward(self, x):
        x = self.encoder(x)
        x = self.decoder(x)
        return x
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model = ConvAutoencoder().to(device)
criterion = nn.MSELoss()
optimizer = optim.Adam(model.parameters(), lr=0.001)

# Assuming the dataset is ready and loaded as TensorDataset or similar
# For example purposes, let's assume we load it using DataLoader
train_loader = DataLoader(dataset, batch_size=64, shuffle=True)

num_epochs = 10
for epoch in range(num_epochs):
    for data in train_loader:
        img, _ = data  # Assuming there are no labels
        img = img.to(device)
        
        # Forward pass
        outputs = model(img)
        loss = criterion(outputs, img)
        
        # Backward and optimize
        optimizer.zero_grad()
        loss.backward()
        optimizer.step()
        
    print(f'Epoch [{epoch+1}/{num_epochs}], Loss: {loss.item():.4f}')

2024-04-08

Artificial Intelligence Data Science and Governance Neural Network

Interleaving algorithm for optimization of neural networks with self-learning perceptrons

by Massimo 2022-01-01

Interleaving is an increasingly popular technique for evaluating information retrieval systems based on implicit user feedback

In one of my research project on data science, I coordinated a new effort to select the best compund from a list of compounds using the interleaving approach, instead of the common A/B Testing,

In particular I used the team draft interleaving, which mimics the process of how team selection occurs for a friendly sports match.

You can find more about this in these links:

https://dl.acm.org/doi/10.1145/2094072.2094078

https://netflixtechblog.com/interleaving-in-online-experiments-at-netflix-a04ee392ec55

This approach can be implemented also in Artificial Intelligence models where perceptrons change their activation functions based on the input data, after a pattern of data is recognized from the perceptron itself or from the control part of the model. In the past, we used standard backpropagation of the error to improve the weights, but we found sometimes perceptrons can change even before the process of learning is completed. This learning during responding has been possible thanking the interleaving approach to choose the best configuration for this model.

2022-01-01

Artificial Intelligence Data Science and Governance Neural Network

The (Un)reliability of Saliency methods – Google Research

by Massimo 2021-07-16

Feature importance evaluation is a fundamental problem in deep model interpretation and feature visualization. Saliency methods is one of the most popular feature importance evaluation methods, assign to each input feature an importance score, which measures the usefulness of that feature for the performance. A large importance score of a feature means a large performance drop when the feature is unavailable. One way to explain deep learning behaviour is the saliency methods. In Google Research page, it is clearly demonstrated that this method does not work.

Indeed, saliency methods aim to explain the predictions of deep neural networks. These methods lack reliability when the explanation is sensitive to factors that do not contribute to the model prediction.
Saliency methods that do not satisfy input invariance result in misleading attribution.

Source: The (Un)reliability of Saliency methods – Google Research

2021-07-16

Artificial Intelligence Data Management Data Science and Governance Neural Network R&D Statistics

Research on Climate Change with statistics methods

by Massimo 2016-08-29

Predicting climate change effects

In the past I supported agrobiodiversity scientists demonstrating the causal effects of climate change on the living of the plants.

Scientists monitor and track more than 15.000 different plant species worldwide.

More than 1.300 Institutes worldwide collect “observations” of plants in a huge portal, the Global Biodiversity Information Facility GBIF https://www.gbif.org/

Today GBIF stores more than 1,3 billion of records catalogued using the data-standard Darwin Core, including observational data (GPS coordinates, altitude, etc.).

Challenges in statistics

How can you find evidence that climate changes is moving plants and/or change their dispersion on the territory. For example, some plants disappeared from where they were observed during ‘60s or appeared in new zones or at higher altitudes. Some other plants spread over mountains because the climate was more favorable due to the increase of temperature. Example: the plant Melampyrum nemorosum which lives in temperate climates has been observed now in Croatia mountains, which has very cold winters.

I collected the observation data needed, described data with descriptive statistics and performed regression analysis to find correlation among plants behavior and changes of the climate.

The Dataset

Datasets I worked on:

– Plants: GBIF portal, 250 million of observations made from 1920 to today

– Climate data from World Meteorological Organization GCOS: I used about 200k records of historical climate data observations

Variables considered:

– 152 variables from plants’ observations (Darwin Core data standard + Observational data)

– 13 variables from climate observations: precipitation, climate air temperature, humidity, wind speed, pressure, lightning, upper-air temperature, carbon dioxide (CO2), ozone, date, location (GPS coordinates)

Features

0 plants' observations

0 climate's observations

Data challenges

Direct plants’ observations made in 60 years were affected by many quality issues:

Human errors: name of the plants wrong, recorded date invalid, date invalid, geodetic datum invalid, elevation non-numeric or swapped, etc.
Technology for observation changed, like the GPS tools (just few units in the 60s, today quite common with smartphones)
Process to create observations changed: Initially scientists catalogued plants in printed forms; then scientists used Word documents, and, only in the last 30/40 years, relational databases have been adopted – I led a project to convert data from thousands printed forms to database of observations with metadata (presentation at TDWG 2010 conference)

Sampling was affected by many (non-sampling) biases. Some examples I remember:

Undercoverage bias:
- Number of observations of plants in a specific area depends on the funds allocated to the Institutes in doing this process. Therefore, the presence of plants in a location depended on the number of observations made in that area. It means only some Institutes had good historical data.
- Scientists observed species only in some areas, probably close to their institutes, and at different times. This means I had to group by time the observations of plants in a map in order to consider which plants live in that area.
Information bias:
1. Each Institute uses its own set of tools and procedures to observe data. This means descriptive statistics are difficult to compare each other
2. Observations happened at different seasons worldwide; in extreme climates (like cold winters or hot months) many plants naturally die so frequency of observations of course change.
3. Climate data refers to the place where is located the meteorological station
Time varying confounders: external factors to the climate change can modify the presence of plants in a zone, like pollution and human activities

Methodology for the problem

For some operations, like interpolation of climate data or calculation of the centroids of plants in a geographical map, I used the methodologies commonly used by agrobiodiversity scientists and climate specialists.

The methodology I followed was tailored for the data, and required in some transformations and assumptions:

GROUPING OF OBSERVATIONS BY DATE: To calculate correctly descriptive measures, I grouped observations of plants by date of observation: all plants were considered “living organism” in the quarter of the year where observation was recorded. So, if a plant was observed in February 1998 then it became part of the group of plants living in the quarter January-March 1998 in that area.
IDENTIFICATION OF CENTRES OF MASS OF SAMPLED PLANTS: I considered the geographical distribution of each sampled plant as concentrated in a single point, the centroid; for each plant sampled, I calculated also its mean and dispersion considering 10 hectares around its centroid. I considered the distance in meters centroids moved in a quarter. The hypothesis was that centroids moved because plants grow differently because of changes of climate factors.
INTERPOLATION OF CLIMATE DATA: For the climate data, since the observations are done in meteorological stations, I performed interpolation of climate data to identify values in the exact location of the plants sampled. I used a deterministic interpolation algorithm called Inverse Distance Weighting (IDW), commonly used by climate scientists.

Unfortunately, I had to consider only those variables having good historical data, in the geographical areas identified after the sampling of the plants: humidity, temperature, pressure, precipitation.

Implementation

A standard process of statistic was followed, considering the quality issues and biases described above:

Multistage samplings of plants observations

Stage 1: Cluster sampling based on Institute names; only some Institutes collected good historical data

Stage 2: Stratified sampling based on qualitative variables of the plants (like kingdom, family name, genus name, taxonRank) – made by biologists

Stage 3: Systematic sampling – this avoided manipulation of the whole frame of Stage 2

The sample size resulted 1.2 million of samples of about 1250 plant species.

Definition of response variables

Variation of GPS position of the centroids in meters (∆centroid) from quarter to quarter: the centroid is the centre of mass of the plants in a map, in a period of a quarter of year. I found the centroids using haversine formula with GPS coordinates
Number of occurrences found in the area of 10 hectares around the centroid
Variation of the Quartile Coefficient of Dispersion (∆QCD): the dispersion of plants in a region of 10 hectares around their centroids were measured by the Quartile Coefficient of Dispersion (QCD). Please note that the computation complexity for the QCD algorithm is excellent: O(n log n), n=number of units observed from a sampled plant

Inverse Distance Weighting (IDW) algorithm

The dataframe was created locally considering the response variables above defined
I accessed the historical observations of climate data from the GCOS institute and applied Inverse Distance Weighting (IDW) algorithm to interpolate the climate data in the exact location of the centroids of the sampled plants
I determined the mean values of the climate data in that point in a quarter period (very time-consuming iterations; today accelerating adaptive inverse distance weighting algorithm could be used)
I created then a dataframe joining data of sampled plants, climate data, observations data, date expressed in quarter/year

Data analysis

Analysis of the dataframe and some findings:
- From time plots of samples, I found seasonality and some trends on number of plants in a region and their spread in that region
- With Polar seasonal plots and scatterplot matrices, I found good correlations of some groups of plants with the climate data: with the help of biologists, we identified group of plants which responded more to variations of climate data. Example: plants that live at high altitudes or plants with good autochthone resilience showed high correlation of dispersion with temperature and humidity
- Forecasts from regression model highlighted that at ground level plants should increase dispersion in their region (∆QCD) and their occurrences in their region should drastically drop
- Some plants found the habitat at different locations: this was evident mostly in dried areas where the ∆centroid values were higher
- Putting altitude of observations as a response variable, I found some plants changed their altitude because of the change of climate in their region. So even if their dispersion in the territory was low (∆QCD) they actually spread vertically over mountains and hills

Getting data

Data accessed and manipulated via dedicated RESTFUL API of the data provider GBIF
Data engineering of local datasets done in dedicated on-premises servers
Development language: python and some R scripts, T-SQL for some data manipulation functions
Languages for data engineering: SQL and stored procedures in Microsoft SQL Server
Geographical tools: ArcGIS, google maps
Development tools: Eclipse

2016-08-29