Yes! By connecting AI to our brain!
Ideas from TED Talks
Can we build AI without losing control over it? (Sam Harris | TEDSummit) https://t.co/tPdyApe8es via @TEDTalks
Yes! By connecting AI to our brain!
Ideas from TED Talks
Can we build AI without losing control over it? (Sam Harris | TEDSummit) https://t.co/tPdyApe8es via @TEDTalks
Building LLM-Powered Agents
I really like how this figure summarizes the key components needed to build LLM-powered applications.
There are a lot of developers already building autonomous systems that achieve complex tasks and LLMs are at the heart of it all.
Prompting LLMs is one thing but you still need to figure out what information to pass to the LLM and how to process and use the information it's returning.
The key components for building LLM-powered agents include:
- Data Sources - for loading data or metadata
- LLMs - leverage models like GPT-4, Claude, or Llama 2
- Code Executor - for executing code and returning results
- Document Retriever - for embedding and retrieving documents
- Other ML Models - for performing other helpful ML tasks like prediction or forecasting
Not all components above are required but combining one or more can lead to all kinds of useful tools like search engines for knowledge bases, chat LLMs on legal documents, and customer support automation.
Learn more here: https://t.co/syJdN2WcFs
Detecting small objects with computer vision models is challenging.
This is because they occupy a few pixels in the entire image.
However, the SAHI technique makes it possible to detect small objects using various models such as YOLOv5.
SAHI is a generic slicing-aided fine-tuning and inference pipeline that you can use on top of any object detector without the requirement to fine-tune the model or change its architecture.
Slicing-aided fine-tuning works by extracting patches from images and making them bigger. At inference, SAHI divides the image into smaller patches and makes predictions from the bigger versions of the patches. The predictions are then converted to the original image coordinates after non-maximum suppression (NMS).
In the video below, I used Supervision which now supports SAHI and YOLOv5 to detect small objects with accelerated inference on CPUs offered by DeepSparse.
Follow @themwiti if you are interested in seeing more computer vision demos.
Links on how to use Supervision in the next tweet.
RLAIF: Scaling Reinforcement Learning from Human Feedback with AI Feedback
paper page: https://t.co/lozxQUDM31
Reinforcement learning from human feedback (RLHF) is effective at aligning large language models (LLMs) to human preferences, but gathering high quality human preference labels is a key bottleneck. We conduct a head-to-head comparison of RLHF vs. RL from AI Feedback (RLAIF) - a technique where preferences are labeled by an off-the-shelf LLM in lieu of humans, and we find that they result in similar improvements. On the task of summarization, human evaluators prefer generations from both RLAIF and RLHF over a baseline supervised fine-tuned model in ~70% of cases. Furthermore, when asked to rate RLAIF vs. RLHF summaries, humans prefer both at equal rates. These results suggest that RLAIF can yield human-level performance, offering a potential solution to the scalability limitations of RLHF.
The traffic analysis project is done! The YouTube tutorial will be out tomorrow. Stay tuned!
Wait till flow counters appear around 0:06.
Github repo: https://t.co/xXMRaS3Guk
Um proprietário de cafeteria usa IA para rastrear o quão produtivos são os baristas e quanto tempo os clientes passam na loja.
Aproveite o seu Frappuccino de Duplo Chocolate enquanto algo está te rastreando. 😂
The traffic analysis project is growing! The YouTube tutorial will be out this week.
Progress: I can now identify that the car is in a specified zone.
Next: Match entrance and exit zones for every tracker ID to analyze the traffic flow.
GitHub repo: https://t.co/xXMRaS3Guk
One of the more interesting computer vision papers I read this week:
They propose applying the Segment Anything Model (SAM) to medical 2D images.
The challenge here is taking a model pretrained on natural images to work on medical images. There is an obvious domain gap.
So the authors proposed a large-scale dataset medical image segmentation dataset containing 4.6M images and 19.7M masks, including various modalities and objects. This is huge for the research community.
The SAM model was then fine-tuned on this dataset and evaluated on medical image segmentation across modalities and anatomical structures. Finetuning is obviously a big deal here and that's shown through the really strong performance obtained on several datasets.
Fun paper to read.
paper: https://t.co/Tla4hinZYW
code: https://t.co/rlaYCoTUOV
YaRN: Efficient Context Window Extension of Large Language Models
paper page: https://t.co/qocIjkanhM
Rotary Position Embeddings (RoPE) have been shown to effectively encode positional information in transformer-based language models. However, these models fail to generalize past the sequence length they were trained on. We present YaRN (Yet another RoPE extensioN method), a compute-efficient method to extend the context window of such models, requiring 10x less tokens and 2.5x less training steps than previous methods. Using YaRN, we show that LLaMA models can effectively utilize and extrapolate to context lengths much longer than their original pre-training would allow, while also surpassing previous the state-of-the-art at context window extension. In addition, we demonstrate that YaRN exhibits the capability to extrapolate beyond the limited context of a fine-tuning dataset.
I am teaching the very first Reinforcement Learning course at National Taiwan University this semester! Dear RL folks, can you describe why students should learn RL in one sentence? Please RT 🙏
A alternativa ao ChatGPT do Google acaba de ser atualizada (disponível no brasil e na europa)
O Bard agora tem alguns recursos que não existem no ChatGPT.
Aqui estão os novos recursos do Google Bard que mudam tudo (disponíveis gratuitamente):
[Thread]
The first (of two) books on postdigital research is now published! https://t.co/xXoquGVeFP
If you want a free pdf, you know the drill - PM your email and it's on its way! :)
Thanks to all authors and co-editors Jeremy Knox and Alison Mackenzie for their great work!