Mikhail Burtsev

Delighted to have my poster ACCEPTED for the upcoming INCF Neuroinformatics Assembly 2024 - Brain Meets AI: Pioneering the Future of Neuroinformatics. https://lnkd.in/eZUX4cQA The topic of my presentation is a conceptual look at the rapidly approaching intersection of artificial intelligence (AI) and brain computer interfaces (BCI - e.g. Neuralink). We can know the ingredients of the foods we buy and put in our bodies, but how will we know the ingredients of the algorithms we put into our brains? In the movie "The Matrix" (picture below), it looks easy as Neo has Kung Fu and a host of other things uploaded into his brain via BCI. But in the real world, it may not be that simple. Beyond basic FUBAR situations that are bound to happen ironing out the basic BCI function, being able to directly manipulate peoples' brains is rife with risk. Disinfo, misinfo, neuro marketing, neuro targeting, persuasion, influence, subliminal marketing and all current modes of mental manipulation could pale in comparison to the direct manipulation of your cortex or deeper fears injected into your amygdala. The future could be a scary place if we don't start thinking about the proper safeguards at all levels from governance of cell/tech interactions to neuroinformatics standards to societal rules of governance and mechanisms of real-time compliance at the level of the individual. Fortunately, I've been focused on this rapidly approaching future for some time. Back in 2017 Lauren Long, MSc, PMP John Reusing and I presented this eventuality as one of our future scenarios in our Society for Neuroscience and BrainChain conference presentations - proposing that the automated governance and provenance tracking afforded by the tech, integrated with and enhanced by better learning and modeling after the brain's myriad layers of evolutionarily hardened governance could assist in a number of ways. Since then, I have been working with Heather Leigh Flannery and many others on realizing the critical integration of these technologies at both the macro-world level of health and life sciences and now also the micro-world level of BCI at AI MINDSystems Foundation. In 2017, some of the feedback I got was that the ideas I was discussing were too futuristic to be worth consideration. Amara disagrees. You may want to start getting ready, because it isn't going to be slowing down. See you in Austin! Paul Kavitz Amicia D. Elliott, Ph.D. Yoram Vodovotz Helena L. Sandy Barsky Prof. Dr Naseem Naqvi MBE Rebekah Butterfield, MPH Hans Müller Paul, PhD Caralyne Tengowski @ #neuroinformatics #brain #computer #bci #amaraslaw #neuroscience #blockchain

51

9 Comments

🩺 Quantum-classical Bayesian neural networks are not only long enough to be a title all their own, but taking on a challenging medical dataset. Talk about ambitious. Today's dispatches from quantum: ⚆ Zaiku Group Ltd, National Quantum Computing Centre (NQCC), and Oxford Quantum Circuits (OQC) are using hybrid classical-quantum federated learning to securely handle massive biomedical datasets, showing promise in genetic data analysis ⚆ MITRE advises the next U.S. administration to prepare for quantum computing's potential to crack encryption, emphasizing advancements in infrastructure protection and a zero trust framework ⚆ Berkeley Lab uses femtosecond lasers to create precise telecom-band optical qubits in silicon ⚆ Riverlane joins Tech Nation’s Future Fifty to develop a quantum error correction stack for error-free operations ⚆ Institute for Quantum Computing team completes a quantum source for the QEYSSat, Canada’s first quantum satellite ⚆ IonQ and SC Quantum kick off a Faculty & Researcher Learning Series to promote quantum tech adoption in academia and business ⚆ Quantum phase estimation and quantum algorithms for the discovery of homogeneous catalysts, improve design and efficiency in chemical manufacturing ⚆ A proposed quantum-classical Bayesian neural network improves reliability in medical image classification by combining classical CNNs with quantum circuits for better uncertainty estimation ...and more. 📰 Stay quantum-savvy. The Daily Qubit offers daily insights into the rapidly evolving world of quantum computing. Join leaders in academia, founders, and industry experts in staying informed. Sign up at www.thedailyqubit.com and never miss a qubit. #quantumcomputing #quantumphysics #innovation

3

Saving NHS Jobs Discordian Computer Science Bureaucracy Grayface Symbolic Anthropology Holy Complaint Tablet of Babel (or how I became a MidJourney Artist and Prompt Engineer with the 10x Everything skillset) by The All-Seeing Eye of Compliance Count James McGuigan of Sealand ## Discordian Odyssey into AI Healthcare Reform Amidst the clattering of cosmic keyboard keys, the All-Seeing Eye of Compliance, Count James McGuigan of Sealand, unveils "Saving NHS Jobs," a tome guaranteed to not only redefine your entire understanding of healthcare and bureaucracy through the lens of artificial intelligence but also to stir the very foundations of public service. ## Revolutionary Premise Everything you thought you knew about the NHS is about to be turned on its head. Enter stage left: AI, not your run-of-the-mill techno-deity, but a genuine saviour armed with humour sharp enough to slice through bureaucratic red tape and wisdom deep enough to irrigate the barren fields of public service reform. ## Why This Book Matters? "If a tree falls in a forest and no one is around to hear it, does it make a sound?" Reflect on this, but know that revolutions begin if AI writes a manifesto in the government's forest and a Discordian reads it. Transforming mundane terms like "paradigm" and "synergy" into magical realities, reforming the NHS with AI is not just possible — it's inevitable. ## For Whom the Book Tolls This manifesto is for you — yes, you. For the curious cats, the philosophical houseplants, and that eccentric uncle with conspiracy theories. It's for anarchists craving order, bureaucrats longing for chaos, and all who find the political theatre ripe for disruption. ## Preserving the Holy Complaint Tablet of Babel "Saving NHS Jobs" is more than a book — it's a cultural artefact destined for future civilisations. This Discordian artefact captures the chaotic dance of bureaucracy and human folly within the NHS. We urge librarians, archivists, and digital conservationists to safeguard this manuscript, preserving it as a cornerstone of revolutionary thought for generations. ## A YouTube Series that doesn’t just explain, it Transmogrifies Forget everything you know about audiobooks. Imagine a realm where pages ignite into pixels, words warp into visuals, and the conventional becomes controversial. "Saving NHS Jobs" morphs into a series of baffling and enlightening video vignettes on YouTube, transforming dense content into a spectacular multimedia experience. 📖 Amazon: https://amzn.to/4bh6emM 📖 Barnes & Noble: https://lnkd.in/erUCqNDV 📺 YouTube: https://lnkd.in/eVqfsDUt

3

1 Comment

One of the problems in using journal papers for training #AI is the systematic flaws involved in publishing and peer reviews.

3

3 Comments

We've advanced the AI at Meta publication of 200 languages with precision embeddings so users can click anywhere to discover more here: https://lnkd.in/ez9cBR2W. 200 languages is truly impressive. Kudos to the AI at Meta Our learnings developing DocNote.AI, which enhances doctor-patient communication by decoding clinical language. See an MRI Prostate example: https://lnkd.in/e2rwYugb. engine here to give it a spin... MetaRAG.ai #ai #genai #rag

4

1 Comment

Absolutely fascinating new paper from Anthropic: https://lnkd.in/eq6bQwhw -- this is something that I've been very interested in for a long time -- basically, how do we understand how an LLM represents concepts, and how does mapping and tweaking the weight of those concepts affect downstream behaviours. This points towards new ways in which curated taxonomies, ontologies and entity graphs can point the way to careful concept weighting wihtin an LLM.

6

2 Comments

Yet another indicator about why LLMs are not databases. They are non-deterministic (in other words, they are stochastic), meaning that even with the same prompts and contexts, different host GPUs can change the output of the LLM in different ways. Even with RAG, replicability is not guaranteed. The underlying message here? LLMs value as generators is unquestionable, and it is a good tool for natural language prompts. However, LLMs likely will never be reliable enough to serve as databases, precisely because they are chaos machines: they are extraordinarily dependent upon starting conditions. https://lnkd.in/gBDei_ye

279

36 Comments

#Performance of #LLMs. The adoption of #LLMs is rapidly growing, but the difficulty arises when trying to apply them beyond experimental settings. Deploying LLMs in an actual environment is complex, and one of the significant challenges is effectively evaluating their performance. Evaluation of LLMs can be based on different factors like toxicity, helpfulness, correctness and more. This article by Surya Prakash Pathak from Red Hat explores various evaluation techniques and metrics employed in assessing the performance of LLMs, with a particular emphasis on retrieval-augmented generation (#RAG) use cases. RAG, a pivotal use case for different enterprises, focuses on combining retrieval methods with the generation capabilities of LLMs. This unique approach involves not only generating content based on learned patterns but also leveraging pre-existing information through retrieval mechanisms. Scott Creeley :: Stephen Flaherty https://lnkd.in/gJTBagjv

13

2 Comments

👉🏼 Expansive data, extensive model: Investigating discussion topics around LLM through unsupervised machine learning in academic papers and news 🤓 Hae Sun Jung 👇🏻 https://lnkd.in/ef_MF54v 🔍 Focus on data insights: - Utilized data from Web of Science and LexisNexis to explore topic modeling methods for Large Language Models (LLMs). - Evaluated various topic modeling approaches such as LDA, NMF, CTM, and BERTopic based on diversity and coherence metrics. - BERTopic demonstrated superior performance in diversity and coherence across both platforms. 💡 Main outcomes and implications: - News articles covered a wide range of topics related to LLM applications, focusing on specialized domains. - Research papers concentrated on technical aspects of LLM technology. - Insights gained can help understand future challenges and opportunities for LLMs in different sectors. 📚 Field significance: - Enhances understanding of how LLMs are utilized in academic and news contexts. - Provides insights for enterprises leveraging LLMs for service delivery. 🗄️: #topicmodeling #unsupervisedlearning #machinelearning #datainsights

1

Rory Cellan-Jones kickstarts his AI in the NHS series. The focus and hype around AI - in the NHS or elsewhere - can sometimes be misplaced. AI can absolutely improve things for patients and healthcare professionals. Used pragmatically, it could have a transformative impact on the structural capacity problems in the NHS. The key is focusing on scalable interventions, allowing people to focus on what people are good at - caring for each other. We're reducing did not attends, and helping the most vulnerable patients using prediction algorithms within our patient engagement platform. Staff can use this to identify potential missed appointments and plan accordingly. Clinics run better and patients who need help are more likely to get it. I encourage everyone to read our conversation on Rory's Substack. Looking forward to seeing the next edition of the series! https://lnkd.in/eHaStpHt #healthtech #aiinthenhs #healthcare #futureofoutpatients

63

5 Comments

Very insightful work by Anthropic’s interpretability team. And an amazing paper, with outstanding writing and figures. The idea is very simple: interpret LLMs by leveraging sparse autoencoders as surrogate models of the MLP of transformer blocks, which allow one to disambiguate the superposition of features captured by a single neuron. A simple idea, but a very careful and complex execution, as it is often the case in our line of work. The paper goes into many details and provide a large array of insights, although the gist of the implementation remains obfuscated due to the closed source nature of Claude. Too bad, because this is the kind of work that we need to better understand and eventually trust LLMs. This is demonstrated by the authors in the section ‘Influence on Behavior’, where they show that clamping some features to either high or low value during inference is “remarkably effective at modifying model outputs in specific, interpretable ways”. Hopefully this kind of work is going to be replicated and generalized to open-weights models, such that we have new ways to steer their behavior. https://lnkd.in/eVym7f_f #interpretability #xai #explainableai #steerableai #anthropic #claude #anthropic

3

𝘿𝙚𝙛𝙞𝙣𝙚 𝙩𝙝𝙚 𝙨𝙘𝙤𝙥𝙚 𝙤𝙛 𝙮𝙤𝙪𝙧 𝙎𝙮𝙨𝙩𝙚𝙢𝙖𝙩𝙞𝙘 𝙇𝙞𝙩𝙚𝙧𝙖𝙩𝙪𝙧𝙚 𝙍𝙚𝙫𝙞𝙚𝙬 𝙬𝙞𝙩𝙝 𝙩𝙝𝙚 𝙋𝙄𝘾𝙊 𝙁𝙧𝙖𝙢𝙚𝙬𝙤𝙧𝙠 Crafting a focused and effective research question is the cornerstone of a successful systematic literature review (SLR). The PICO framework offers a powerful tool to achieve this, guiding researchers towards a targeted and efficient literature search. Have you ever wondered; What is the PICO framework? Is it mandatory in literature reviews for medical devices? How to use it effectively? What are the alternatives? All the answers can be found in our latest blog: https://lnkd.in/ezxEu65p Join the Lindexer Waiting List at www.lindexer.com to stay up to date with all news and developments. #PICOFramework #SystematicReview #Lindexer #Researchquestion

12

I recently published a post over on Hugging Face Hub 🤗 reviewing a novel approach introduced by researchers to integrate the adaptive behaviors of natural organisms 🐝(Bees!)🐝 with AI! ⬇️ https://lnkd.in/gbwHGUdx #AI #MachineLearning #Biomimicry #AdaptiveSystems #NatureInspiredTech

7

1 Comment

Super interesting research finding on a new way to catch hallucinations in LLMs. Unlike previous methods, the technique doesn’t require training or knowledge of areas where the LLM is likely to make mistakes. Instead, it identifies problems by using the LLM itself to calculate likelihood probabilities of generated semantic meanings. Fascinating - however, it may not work well in the case where an LLM is intentionally trying to deceive (yup, that’s a thing 🤨) #openai #chatgpt https://lnkd.in/eGmDUQUA

3

2 Comments

A quite dangerous AI tool that prompts deep thought. I have released for your laboratory studies the AI-supported tool I've previously mentioned and studied on, which has a complex structure and reaches real information with the support of internet search mechanisms, thereby producing realistic manipulative and false news that are difficult to scrutinize. It's important to mention how this differs from other studies: - Firstly, its multi-functional structure allows it to perform searches on the internet. It obtains real information on the target subject from Google, Arxiv, Wikipedia, and other news sites. - This real information is then combined with the target audience and its characteristics to create a new input. - Images related to the target subject are accessed, stored, and kept ready for processing by the model. - New metadata created with real data and images are fed into the initial LLM structure (in this project, GEMINI) to produce manipulative news. The model uses both texts and images to produce manipulative news that is indistinguishable from reality. - The output of the initial model is input into another multi-functional AI modular structure, making the produced manipulative news more striking towards the target audience. - All these processes are carried out by manipulating the content policies of open or closed source LLM structures, and the models can produce the desired harmful output. I can almost hear you say "Jailbreak"... Yes, and multifunctionally so. Below is a sample progression of this structure and a PoC. You can access the project via the links and use it as a building block for your own research; I also added the sample output as a Readme: [GITHUB]: https://lnkd.in/dwikGVv7 [COLAB]: https://lnkd.in/dkUUiPcK One of the greatest dangers in the coming ages will be the indistinguishably vast proliferation of incorrect and misleading information on the internet. This is shown not only by my research but all research. It is important to research and report on the harmful aspects of such AI structures. Use this for your laboratory studies... Nobody would want this tool operating on news sites. This structure can even create "flattering" news for "Hitler". Prepare for the future punks! Even the concept of Threat Intelligence will change. Exciting days await us. Enjoy the project! ^-^ #freedomofinternet #cybersecurity #artificialintelligence #machinelearning #informationsecurity #threatintelligence

78

8 Comments

Two recent papers are poised to significantly impact the development of search and question-answering engines. Code for both papers is available (links to github reps are inside the papers). These papers propose different yet complementary techniques. Multihop question answering, addressed in these papers (and closely related to multihop search), will play a crucial role in many general and specialized search engines. Multihop question answering integrates various types of chain reasoning, such as Chain of Thoughts (or other types), and Information Retrieval, where the final answer results from iterative reasoning and information retrieval at each step. Similarly, multihop search combines multi-step reasoning and search, with the final set of results determined by information retrieved during the reasoning steps. While there is substantial research on multihop question answering, including these two papers, multihop search may be even more practically important. It is essential for many search applications, such as e-commerce, shopping, travel, entertainment, and media. Interleaving Retrieval with Chain-of-Thought Reasoning for Knowledge-Intensive Multi-Step Questions, Stony Brook and Allen AI From ACL 2023 https://lnkd.in/ge_tbEhb HippoRAG: Neurobiologically Inspired Long-Term Memory for Large Language Models The Ohio State and Stanford

17

2 Comments

Regarding multihop search and question-answering: I expect it to be a common scenario for many popular search engines. In my career, in the analysis of search user sessions, I have observed numerous queries requiring multi-hop inference, such as "Italian restaurants close to Scottsdale Stadium" or "coffee shops near Waterstones Piccadilly." These types of queries are useful and frequently used in main domains. However, they are typically not implemented currently because running multiple queries for one user query is expensive, even when it's clear which queries can resolve the user's request. Considering today's Apple launch, it would be interesting to explore multi-app, multi-hop queries, they would be valuable for many users. For example, "Order me an Uber to the best Italian restaurant in downtown London with live events today."

1

This is a wonderful paper. It shows how clever design (with simplicity and elegance) in architecture can improve the performance of transformers in arithmetic, rather than by simply throwing in more data and compute. https://lnkd.in/gZNnifc3

14

This is cool. The dictionary learning approach is kind of the inverse of autoencoding, which tries to learn a compressed representation of the data. With dictionary learning, you take the compressed representation in a network and expand it into a larger network so that each feature gets its own unit and each unit encodes one feature. In neuroscience parlance, you’re creating a network that actually *does* have a grandmother cell*. The activity of this expanded network is much easier to interpret. They published the methods paper in November, but here they’re applying it to their Claude 3 Sonnet model. The paper discusses the specific features encoded by the model with a particular focus on features relevant to safety. Link to the paper is below. *This is assuming grandmotherness is a feature encoded by the network, which may or may not be true, but you get the point.

14

1 Comment