Alpes has the fastest learning algorithm with O(NlogN) time complexity. This will allow you to iterate training with your data fast to find the right feature set.
Algorithm can perform incremental learning. When fed with new training data you don't need to run the whole training process again. This will allow your products to imbibe new data on the fly while making decisions.
Clear understanding of why the system is learning or not learning. AI is not a blackbox anymore. This will allow you to fine tune your data and get the right data to make your product learn better.
Well defined generic train and test APIs for all kinds of data. Specific APIs for speech, vision and text problems. Easily integrate into your developer stack.
Host the trained models yourself so that your data never leaves you. Or run the trained models from the Alpes cloud.
Enable voice and conversational intelligence into your products
Apply a natural language layer to your products
Integrate Computer Vision into your existing product
Power your shopping cart recommendation engine with Alpes NN.
Text classification, summarization and sentiment analysis have all given good results.
Disease prediction has worked well on public datasets.
We are currently in beta mode and access to API is invite only. Please signup here
Extract features from your data. Call the upload API and then the train API. You will have your trained model ready in minutes.
Use our test or predict API calls for testing or to make a prediction on your data.
You can download the trained model and use it in your application or you can directly run the model from our servers.
The above research work on Neural Networks, stems from our discovery of an algorithm that can separate N points in d-dimensional space by hyper-planes, in such a manner that all points are separated from one another by hyperplanes. This algorithm leads to a very important breakthrough in neural network research. Because it enables one to train a neural network and arrive at the configuration of the neural network and determine all the weights of the processing elements in a non-iterative manner 1. It takes only an order of NlogN multiplications, where N is the number of points (samples) in d-dimensional space. 2. The number of planes q needed to separate N points in d-dimensional space is (only) approx. q=log2(N); this is especially true when the dimension d of the space is large: N < 2^d 3. It is non- iterative and has to terminate successfully. For example: 50,000 random points in a 25 dimensional unit sphere can be separated from one another by only 27 planes.
A) Mathematical Basis: Most importantly , these new algorithms have a mathematical basis and water-tight proofs which guarantee their success whenever and wherever patterns are recognizable by given features and the algorithms are deterministic and non-iterative. The algorithm determines the architecture. B) Accuracy: Whatever problems can be solved by the conventional Neural Network (NN) or Deep Learning technique can be solved by these new algorithms. Several particular cases have already been taken and solved to reinforce this view- viz (i) digit recognition MNIST, (ii) Face-recognition viz. Extended Yale data base, and (iii) Alphabets. C) Time: The time complexity of the algorithm is NlogN and is as fast as the FFT. For instance when we apply our method to the Extended Yale database consisting of approximately 11,000 face images each consisting of (30 X 30 pixels)i.e., 900 dimensions. The time taken to separate each data point (image) was only 40 milliseconds in a Laptop (21 planes were needed); and the time taken to classify 3700 test images was 2.5 minutes. By the conventional Back-propagation method the neural network took almost 4 days of training on the same machine. D) Restart Feature: This method does not need to restart when new data is found during the learning process (or later). The algorithm can start learning from where it has left off. It also makes optimal use of information (as described above). One may contrast the above situation with the usual Back Propagation algorithm (discovered by Rumelhart, Hinton and Williams in 1986 and used in every NN and Deep Learning Application), where there is no proof and no guarantee of success and is also iterative and non-deterministic and even the initial architecture needs be guessed, and therefore solely based on trial and error.
CEO, M.S. (IIT Kanpur), Ph.D (University of Madras)
Vice President, Ph.D University of Cambridge
We at Alpes are striving hard, to Apply our research for real-world impact in health, science, energy and many more domains. Alpes is one of the world leaders in artificial intelligence research and its application for positive impact. Our motivation in all we do is to maximise the positive and transformative impact of AI. We believe that AI should ultimately belong to the world, in order to benefit the many and not the few, and we’ll continue to research, publish and implement our work to that end.