Hello readers! In this issue, we cover

• Model collapse: when models are trained on data generated by LLMs, they tend to degrade in performance over time

• Amazon improves image classification with multiple choice questions

• AI can predict stock prices using sentiment analysis from news sources

• AI can discover HIV-inhibiting molecules

• Using LLMs for medical image analysis is still not completely trusted

🏙️ Amazon Improves Few-Shot Image Classification with Multiple Choice Questions

Amazon researchers improve image recognition when only a few examples are available by using multiple-choice questions to gather extra details about the images. A model answers these questions, enriching the image's information, and then this detailed data is used to classify the images more accurately. This method is particularly useful for distinguishing fine details, like different clothing styles. The new approach outperforms traditional methods in these tasks.

Read more

🔃 Model Collapse: What happens when AI models train on AI-generated data?

The use of LLMs at scale to publish content on the Internet will pollute the collection of data to train their successors.

Research published in Nature explores what happens to GPT-[n] models once LLMs contribute much of the text online. They found that indiscriminate use of model-generated content in training causes irreversible defects in resulting models. They coined the term model collapse, which is a degenerative process whereby, over time, models forget the true understanding of the data.

Read more

📈 Researchers use AI to predict stock prices

Researchers predicted stock prices using a deep learning model called FinBERT-LSTM, which combines sentiment analysis of financial news with historical stock data. FinBERT, a pre-trained Natural Language Processing model, is enhanced with Long Short Term Memory (LSTM) architecture to analyze news categories such as market, industry, and specific stocks. The model was tested on NASDAQ-100 data and Benzinga news, showing that FinBERT-LSTM outperforms other models like LSTM and DNN in prediction accuracy.

Read more

🧪 How AI discovers new molecules to combat HIV

Researchers created the Diff4VS model, a Classifier Guidance Diffusion model with a screening strategy, that can discover potential HIV-inhibiting molecules for the first time.

They also observed a new phenomenon, called Degradation, when using molecule generation models. Compared to real molecules, the generated molecules have a lower proportion that is highly similar to known drug molecules.

Read more

🩻 Are LLMs Trustworthy When Analyzing Medical Images?

In a recent comprehensive study on foundation models in medical image analysis, researchers highlight significant trustworthiness issues associated with these models. Foundation models show promise in advancing tasks such as segmentation, report generation, medical Q&A, and disease diagnosis. However, concerns regarding privacy, robustness, reliability, explainability, and fairness persist.

By reviewing current literature, the study maps the strengths and weaknesses of foundation models in diverse applications and underscores the importance of addressing these challenges to fully realize their potential in healthcare.

Read more

🤯 Today I learned

Every issue, we highlight new AI concepts and terminology to help educate our readers. This issue we learned about:

Model Collapse

The degradation of a model’s performance over time, particularly when new data diverges significantly from the data on which the model was initially trained. Causes of model collapse include:

1. Training on AI generated over time i.e. the output for one model becomes the input for another

2. Training on low quality data

3. Overfitting i.e. when a model learns its training data too well

4. Lack of model robustness

5. Data drift

LSTM

LSTM (Long Short-Term Memory) is a type of neural network designed to handle sequences of data and remember information for long periods, making it ideal for tasks like language modeling and time-series prediction. It uses special units called memory cells and gates (input, forget, and output gates) to control the flow of information and maintain context. This architecture helps solve the vanishing gradient problem, allowing the model to learn long-term dependencies effectively.

Visual Q&A Model

A visual question answering (VQA) model is an AI system that combines computer vision and natural language processing to answer questions about images. It analyzes the visual content of an image, interprets a text-based question, and generates a relevant answer. VQA models have applications in accessibility, image search, robotics, and education. They represent a step towards more sophisticated multimodal AI systems that can process and integrate different types of data.