March 2023 Summaries

Adversarial examples are slight alterations to input data that cause AI models to produce incorrect outputs, often unnoticeable to humans but vast to ML models. These adversarial attacks can be problematic for safety-critical applications like self-driving vehicles and cancer screening. Researchers are investigating ways to make AI less vulnerable to these attacks by detecting, understanding, and defending against them. Some approaches include data poisoning and creating physical objects that resist object detection algorithms or facial recognition systems.

Neural networks (NNs) have become a ubiquitous part of modern technology, underpinning systems like Google Translate and TikTok by functioning as the "brains" of machines, enabling artificial intelligence. The concept can be understood through a simplified model of a speech recognition application, where NNs process input waveforms and convert them into strings by breaking down the audio into chunks and using a layered structure to recognize increasingly complex audio features. Each layer in the NN builds upon the previous one, moving from frequency chunks to letter sounds, then to syllables, and finally to words, with the output layer providing a probability distribution for the possible words. While this model offers a mental image of NN functionality, the detailed mathematical operations and training processes involved in real NNs are more complex and will be explored in future discussions.

Chatbots and conversational AI are improving accessibility for people with disabilities by providing assistance with daily tasks, communication, and mental health support. Virtual assistants like Siri and Amazon's Alexa have enabled individuals with physical disabilities to control smart appliances and interact more easily with technology. Conversational AI technologies also offer benefits in speech therapy, memory aids, and early detection of Alzheimer's Disease. Microsoft-funded projects are exploring the use of chatbots in finding employment for people with disabilities and providing assistance in the workplace. However, challenges remain, including limited research on long-term mental health support provided by chatbots and issues related to voice recognition for individuals with speech disorders. Additionally, hand-gestural languages like American Sign Language have yet to be fully integrated into conversational AI technologies.

Crafting AI Commands: The Art of Prompt Engineering discusses the growing field of Generative AI and its potential to revolutionize various industries. With advancements in models like GANs, VAEs, and Transformers, Generative AI has enabled the creation of original outputs by learning from large datasets. One key aspect of these models is their ability to generate responses based on text inputs or "prompts." Prompt engineering involves iteratively figuring out how to best instruct a model for a desired output. This skill is becoming increasingly valuable as companies look to leverage LLMs and integrate them into their products. While the emergence of prompt engineering raises concerns about job displacement due to AI innovation, history shows that technological advancements often create new categories of work and increase global GDP.

The week of March 12-17th saw significant advancements in AI technology with the release of Stanford's Alpaca 7B, OpenAI's GPT-4, MidJourney's V5, and PyTorch 2.0. These developments have expanded accessibility to large language models (LLMs), improved multi-modal processing capabilities, enhanced diffusion models for image generation, and revolutionized deep learning training and deployment. Additionally, Google and Microsoft announced integrations of generative AI into their consumer products, Workspaces and Copilot respectively. The rapid pace of innovation in the field is expected to continue shaping the future landscape of artificial intelligence.

Deepgram is a foundational AI company whose primary goal is to discover, develop, and apply frontier technologies in machine learning to the real world. Their main product is a Speech-to-Text API that utilizes deep neural networks. The origins of Deepgram can be traced back to their early use of PyTorch as a framework for handling massive amounts of intense calculations efficiently. PyTorch has been especially useful for automated and manual model training, allowing the company to develop impressive ASR and NLU models.

Deepgram has released a new feature called Endpointing, designed to provide customers with transcriptions as quickly as possible after speech detection. The Voice Activity Detection algorithm identifies pauses and notifies users when someone finishes speaking, finalizing the results for the processed time range and returning the transcript with a speech_final parameter set to true. This feature is beneficial for businesses that need real-time transcriptions and can help detect different speakers and analyze customer interactions. Deepgram continues to innovate in speech-to-text transcription technology, offering users an edge in gaining insights from spoken data.

Word Vectorization is a technique used by large language models (LLMs) to learn how to write like humans. It involves transforming words into numbers, or vectors, which represent the relationships between words based on their frequency of co-appearance in documents. These word vectors can be manipulated using mathematical operations such as addition and subtraction, allowing LLMs to understand context and complete sentences. Models like BERT and GPT-3 utilize this technique to generate human-like text by processing large amounts of data during training.

Deepgram celebrates International Women's Day by embracing equity and promoting the inclusion of women in tech and voice technology. The company aims to cultivate and amplify voices that were previously unheard, empowering women at Deepgram to own important discussions and promote representation of cis and trans women voices in speech data. DG Ladies, an internal Women in Tech employee resource group, offers mentorship, skill-building opportunities, and fosters connections between women across the company. The work towards achieving equity is ongoing, with a focus on building a culture where all women can lead conversations without questions or caveats.

Designing conversational AI involves understanding the structure of sentences and creating a flow chart for interactions. Key terminology includes wake word, utterance, intent, slot, and prompt. When designing a bot, consider modality (text or voice), personality, and accents to ensure users feel heard and understood. A successful interaction script can serve as a base for building the conversational flow. Prototyping tools like Figma or Voiceflow can be used to design the flow chart. Keep prompts small and use clear language to make it easy for users to know what to say. Cover corner cases by anticipating potential errors and designing solutions. Design for real people and conversations, considering their environment and needs. Multimodal communication, such as combining voice and screen interfaces, can enhance the user experience. Analyzing conversations through sentiment analysis, topic detection, and summaries can help improve chatbot performance.

The National Federation of the Blind filed a lawsuit against Target Corporation for making their website inaccessible to people with visual impairments, leading to a $6 million settlement fee and improved web accessibility features. Alt texts were originally used to describe images that took long time to load on 1990s dialup but today they are an essential accessibility feature for the internet. As web accessibility gains importance, many companies are now making accessibility a priority by adding alt texts to improve photo descriptions for visually impaired people. The abundance of labeled images with alt text pairs can be compiled into vast datasets and used to train AI models, reducing costs and improving scalability.

The article discusses the top 5 machine learning frameworks for 2023, including PyTorch, TensorFlow, Matlab ML framework, Keras, and Scikit-learn. These frameworks are designed to make neural networks and AI algorithms easy to use and provide a wide range of features for data preprocessing, feature engineering, model training, and evaluation. They also offer various algorithms for supervised and unsupervised learning, as well as deep learning. The article highlights the benefits of each framework and provides examples of companies that use them.