- Letter from the CEO
- Development Update
- Research Update
- NTK Update
Letter from the CEO
“Imagination is more important than knowledge. For knowledge is limited, whereas imagination embraces the entire world, stimulating progress, giving birth to evolution.” — Albert Einstein, What Life Means to Einstein (1929)
Today’s Artificial Intelligence (AI) technologies work by inferring the relationship between input data (e.g. images captured by a camera) and output labels (presence of a cat in the image) by using tremendous amounts of well-labeled data to ‘train’ neural networks. This process is inherently gated by the amount and quality of data. The mantra in Silicon Valley for the last decade has been that “one who has the most data wins” and major platform companies like Google have built data silos to maintain their competitive advantage.
Neuromation began by questioning this fundamental mantra. Can we use synthetic data, or digitally created data that mimics sensory input, to augment or replace real-data to train AI models? If we can simulate the real-world, what new things can we do? If we are unbound by the constraints of acquiring and labeling real-world data, how far can our imaginations take us? If our minds can grasp and learn from imaginary data (cartoons, films, etc), why can’t AI learn from these abstract representations of the real-world?
We were among the first to see the value of synthetic data and we are now seeing companies like Nvidia and Apple using synthetic data together with generative networks to improve tumor detection and enable better gaze estimation, respectively. Neuromation continues to be a thought-leader in the space and in the last quarter we began a focused research effort to answer key questions regarding the potential depth and breadth of applications enabled by synthetic data and hybrid architectures (real + synthetic data). We have made great progress and will be releasing a white-paper in the coming months. If (when) we are able to show the generalizability of synthetic data, we expect to disrupt how AI models are developed leading to democratization of AI technologies.
In parallel with our internal research efforts, the product development team has made great strides in building a best-in-class platform focused on developer experience. We are using the platform internally to support our research efforts and seeing significant improvements in developer productivity. The platform is synergistic with our synthetic data research efforts and together will lead to a material reduction in the time and cost to develop world-class models.
We are excited to begin an Early Adopters Program in the coming quarter to put our tools in the hands of AI researchers and developers. In this report, we will lay out the development roadmap and provide a detailed description of our approach to optimizing the model development lifecycle.
Our team thanks you for your continued support and we look forward to closing out the year with some significant news and developments.
Neuromation is pleased to provide our Quarterly Report, for Q3 2018. The purpose of this report is to provide NTK holders, supporters and other community members with ongoing information regarding the progress at Neuromation.
In our last quarterly report, we made several important announcements, including introducing our new CEO and CTO, restating the company’s Roadmap, and formally presenting our newly defined corporate mission, vision and strategy.
This quarter, we are providing deeper visibility into our product release schedule and the functionality of the Neuromation Platform and are sharing the progress achieved to date by our development team. We also provide updates on accomplishments of our Research Department and other relevant company activities and developments over the past quarter and provide context on the performance of the NTK token in the quarter.
The Big Picture
The Neuromation development Roadmap can be described in three distinct phases:
As shown in the graphic above, there are 3 phases to Neuromation’s near-term product development plans. Phase 1, which was completed in the quarter ending on September 30, was devoted to defining the product and commencing development.
By the end of last quarter, Neuromation successfully completed testing of the base hypotheses stated in the Neuromation Whitepaper. Core architectural principles and design of the Platform were laid out and work on development of the Platform and associated toolsets had already begun.
Phase 2 (currently underway) involves continued Platform testing with select community members, focused specifically on AI professionals. We will establish an Early Adopters Program in Q4, 2018 and grow it to a broader community in Q1, 2019.
Phase 3, which is scheduled to begin in roughly six months, will involve launching the Platform to a much wider audience including all AI professionals, practitioners and enthusiasts. Platform marketplace functionality will continue to expand during this phase helping establish Neuromation as the preferred destination for applied AI developers.
The Life Cycle of an AI Model
The completion of Phase 1 of the Development Roadmap resulted in several conclusions. One of these was the establishment of a workflow model for a typical AI research project. This is being used to guide product decisions and assess development progress.
We see AI research workflow as a staged process spanning from model inception and set-up, to small-scale model training, and on to large-scale model training and deployment. At completion, the Platform will address all 3 stages of model development.
- Stage 1, the AI Developer will use the Platform to pull the most relevant model to a local environment and make initial adjustments for the tasks at hand.
- Stage 2, the AI Developer will push their altered model to a cloud-hosted interactive GPU environment, where they will use a subset of selected data (10% for example) for initial testing of the performance of the model. This stage allows the Developer to run short iterations with regular alterations, to debug their model, and to prepare for large-scale testing and training.
- Stage 3 will allow the Developer to train their model on the entire dataset, undertaking full-scale testing of multiple models and parameters using large numbers of GPUs.
Accomplishments in Q3
Neuromation’s approach to Platform development is based on internal testing, refinement and adjustment through daily, “hands-on” collaboration by our Research and Engineering teams. Our Research and Engineering teams are focused on two types of jobs: these include the testing of base functionality for a broad range of AI Developer needs encountered during ongoing partner and client projects.
As of this writing, Stage 3 functionality is materially complete and within weeks, roughly 80% of Neuromation’s development work will be performed exclusively on the Platform.
The next key feature to be added will be the implementation of a remote shell and interactive debugging on the Platform that should provide for continuous model tuning on smaller subsets of initial datasets.
Limited Alpha testing is scheduled to begin in Q4, the current quarter, and will initially include Stage 3 of the workflow described above, while Stages 1, 2 and 3 will continue to be developed and released over time. In the following quarter, limited Beta testing will commence, where the goal is for Stages 1, 2 and 3 to be tested by the entire Beta test audience.
Additionally, we should note that it will always to be our priority to remain receptive and flexible to market demand and not only deadlines — so as broader Beta testing commences and the product is being used by a wider audience, we may be adjusting our feature sets as well as overall deadlines.
In Q3 2018, the Neuromation Research Team has focused work on three key areas. The first is core research on synthetic data and integration of research workflow into the Platform (as described above). The second area is fundamental research in collaboration with academia. The third is applied research for specific industry projects. There are a number of projects nearing completion or in progress that we will disclose over the coming weeks. Below is an overview of projects completed during the quarter or currently in progress.
- Synthetic Data — as a core feature of the Neuromation offering, we have devoted substantial resources to research initiatives centered around the generation and use of synthetic data for both computer vision and other applications. We are currently preparing a whitepaper on this subject that will present our research results and stance on the viability of synthetic data.
- We have developed a novel deep learning model for the identification of medical concepts in social media posts that advances state of the art in the field of natural language processing (NLP). Our paper detailing these results has been published in the Journal of Biomedical Informatics, one of the most respected outlets for this type of research. Please refer to our blog post explaining the results and context of this research. Congratulations to Neuromation NLP gurus Elena Tutubalina and Zulfat Miftakhutdinov!
- We successfully participated in the GIANA 2018 Endoscopic Vision challenge, a medical imaging competition at the 21st International Conference on Medical Image Computing and Computer Assisted Intervention (MICCAI 2018). Alexander Rakhlin, a Neuromation researcher, led this project and we are expecting publication of our model for the Endoscopic Vision challenge to follow shortly.
- We continue our close collaboration with Insilico Medicine. In particular, our new paper on natural language processing was accepted by a top research conference, ACML 2018, where we will present a new approach to morphological agreement and apply it to conversational models. In addition to this, we have dedicated substantial resources to a large paper focused on the biomedical sphere, and it is currently reaching the final stages of completion.
- Neuromation participated in Basel Life 2018, one of the largest events on healthcare in Europe; Neuromation CRO Sergey Nikolenko presented Neuromation’s efforts in healthcare at the congress. Representatives of large global pharmaceutical and medical companies were in attendance, and we made many new contacts for future collaborations. Please see our report on the conference.
6. Our joint research project with the European Molecular Biology Lab (EMBL) is advancing well; we have developed a novel model for learning latent representations of single-cell imaging mass-spectrometry data. Results of this research are currently being prepared for publication.
7. Neuromation also advanced our joint project with SolidOpinion, a natural language processing initiative centered on extracting useful data from user comments mined by SolidOpinion from top media outlets on the web. We have developed an engine that combines leading edge advances in named-entity recognition, named-entity linking, sentiment analysis, and other NLP tasks. We have prepared our results for publication and submitted the paper to a major conference.
The Research Team continues its work on fundamental and applied research, and looks forward to a highly productive Q4 of 2018 and Q1 of 2019.
New publications with Neuromation affiliation
- D. Polykovskiy, D. Soloviev, S.I. Nikolenko. Concorde: Morphological Agreement in Conversational Models. Accepted to 10th Asian Conference on Machine Learning (ACML 2018), 2018.
- Vitalii Demianiuk, Sergey Gorinsky, Sergey Nikolenko and Kirill Kogan. Distributed Counting along Lossy Paths without Feedback. Accepted to 25th International Colloquium on Structural Information and Communication Complexity (SIROCCO 2018), 2018.
Global crypto markets continued their selloff in Q3,creating pressure and volatility across all crypto asset classes. ETH fell the most of the major cryp tocurrencies during the quarter, falling from 18% of total cryptocurrency market cap to 11% by the end of the quarter and 10% by this writing. Against this backdrop, BTC emerged as the safest asset in the crypto space, with its percentage of total cryptocurrency market capitalization growing from 43% to over 50% by the end of the quarter and now reaching 53% as of this writing. In dollar terms, ETH declined by 47% during the quarter and over 50% when compared to Bitcoin. This decline has continued since the end of the quarter.
This environment created significant pressure on the token market with few tokens spared from declining prices and volumes. The price of NTK decreased during the period from June 30 to September 30 by 62% in dollar terms and 29% in ETH terms. A positive price move since the end of the quarter has now left NTK 50% below where it started Q3 in dollar terms but with a gain of 7.3% in ETH terms as of October 24, the last day before this report goes to press.
The fall in ETH may have had particular impact on ERC20 tokens, especially for companies with utility tokens that are still in the process of building out the economies on which their tokens will be used. As such, we believe that until such time as NTK is being widely used on the Neuromation Platform, NTK volatility will likely be highly dependent on moves in ETH.
As described in the Q3 Development Update above, our current development timeline foresees Beta testing of the Neuromation Platform to begin in three months, at which time select users will already be actively using NTK to perform AI development. General availability of the Platform is scheduled for approximately 6 months. At this point, we will begin to see the rollout of the Neuromation Platform to the broader public and should begin to see increased use of NTK on the Platform.
Neuromation remains committed to our token holders and pledges to use the Neurotoken (NTK) as the exclusive means of conducting transactions on the Neuromation Platform going forward.
We should also reiterate that all Neuromation revenue from AI development work will be transacted in NTK on the Platform through 2019. Furthermore, Neuromation pledges to always consider the impact on token holders of all new business ventures or product offerings.
The high volatility of the crypto markets of Q2 and Q3 2018 has also made clear the necessity of having robust mechanisms in place on the marketplace to insulate buyers and sellers from currency fluctuations when entering transactions. We have developed multiple potential approaches to this issue that we will be reviewing and testing as part of our development roadmap. These solutions include seamless gateways to enter and exit NTK from fiat or cryptocurrencies.
Neuromation believes strongly in transparency and communication with our token holder community and will continue to report on our progress in terms of development and research activities on a quarterly basis. To this end, Neuromation will be activating English and Japanese Telegram channels and a Chinese WeChat channel for real-time communication with our token holders. We look forward to hearing from you in the future and sharing with you the results of our upcoming Alpha and Beta tests and the exciting release of the Neuromation Platform next year.