The Adversarial Robustness Toolbox (ART) = framework that uses generative adversarial neural networks (GANs) to protect deep learning models from security attacks
Thread⬇️
GANs = the most popular form of generative models.
GAN-based attacks:
+White Box Attacks: The adversary has access to the training environment, knowledge of the training algorithm
+Black Box Attacks: The adversary has no additional knowledge
2/⬇️
The goal of ART = to provide a framework to evaluate the robustness of a neural network.
The current version of ART focuses on four types of adversarial attacks:
+evasion
+inference
+extraction
+poisoning
3/⬇️
ART is a generic Python library. It provides native integration with several deep learning frameworks such as @TensorFlow, @PyTorch, #Keras, @ApacheMXNet
If you'd like to find a concentrated coverage of ART, click the link below. You'll move to TheSequence Edge#7, our educational newsletter. thesequence.substack.com/p/edge7 5/5
• • •
Missing some Tweet in this thread? You can try to
force a refresh
Let's dive into one of the newest concept of synthetic data generation -
active inheritance.
Proposed by @CohereForAI, it's a strategy used in ML to intentionally design synthetic data to achieve specific goals.
Here's how active inheritance works:
1. What's in the base?
The base is a knowledge distillation technique, where a smaller LLM (student) learns from a larger, more powerful model (teacher).
The student tries to mimic the teacher outputs for the same input prompts by learning from the data the teacher generates.
2. Measuring data characteristics:
Cohere researchers use specific functions to track length, diversity, and toxicity in outputs from both teacher and student LLMs. This helps them understand what properties the student inherits from the teacher.
A new model, Meteor, leverages multifaceted information and a Mamba architecture to enhance comprehension and response capabilities in vision-language tasks.
Let's explore its architecture and training strategy👇
1. Meteor's architecture includes:
- a vision encoder (CLIP-L/14)
- vision and tor projectors (MLP modules with GELU activation)
- the Mamba-130M architecture for computational efficiency
- the InternLM2-7B as the backbone LLM.
2. Visual instruction tuning dataset
Covers a wide range of capabilities, including image understanding, common-sense knowledge, cognitive reasoning, and math. It contains 2.1 million question-answer pairs from various sources and diverse real-world images.
TimeGPT is the first foundation model specifically designed for time series analysis.
It excels at generating precise forecasts across a diverse range of datasets and domains.
Here's what you need to know about it:
1/8
The model leverages a Transformer-based architecture, optimized for time series data, with self-attention mechanisms that facilitate the handling of temporal dependencies and patterns across varied frequencies and characteristics.
2/8
It incorporates an encoder-decoder structure, local positional encoding, and a linear output layer designed to map decoder outputs to forecast dimensions.
3/8
1. @cohere LLM University 2. @huggingface NLP course 3. @databricks courses
and more!
🧵
1. @cohere LLM University
The course offers insights into how LLMs work, and their practical applications, and guides participants on using LLMs to build and deploy applications.docs.cohere.com/docs/llmu
2. @huggingface NLP course
This course provides comprehensive knowledge about Hugging Face Transformers, Datasets, Tokenizers, and the Accelerate tool in the field of Natural Language Processing (NLP).huggingface.co/learn/nlp-cour…
DoRA (Weight-Decomposed Low-Rank Adaptation) sets a new standard for optimizing AI models.
It combines the benefits of full model fine-tuning and LoRA.
How does it do that? Let's see 👇🏼
1/7
The genius of DoRA lies in its unique handling of pre-trained weights.
It separates these weights into two parts:
1. one that determines the size (magnitude) 2. one that determines the orientation (direction) of the weight vectors
2/7
This allows for more nuanced adjustments during fine-tuning, closely mimicking the comprehensive learning achieved through full model fine-tuning but with far fewer parameters needing adjustment.
3/7
Want to understand foundation models, generative AI models, and transformers?
Here is your FREE list of 15+ resources to do that:
1. Efficient Transformers: A Survey explores the evolution of Transformer models in various domains. It provides a comprehensive overview of different Transformer variants (X-formers) to guide researchers. arxiv.org/abs/2009.06732
2. A Survey of Transformers
The survey provides a comprehensive review of Transformer variantsin AI fields. It introduces the vanilla Transformer and a new taxonomy of X-formers, covering architectural modifications, pre-training, and applications. arxiv.org/abs/2106.04554