Adkodas: Our Services
Whether you are confounded by all the buzz words: data analysis, big data, artificial intelligence, machine learning, deep learning, analytics... or think you are missing out, because you have data and have heard all the noise but don't know where to start -- contact us. Adkodas advice is a free service.
You have a general idea of the kinds of predictions or patterns you would like to see from your data but want to understand how the most important predictions and patterns are found? Adkodas rule extraction is for you.
Have data and want to project detailed results of new instances that could fall into the same data set? (New customers - new trends - new science - new business?) Rule extraction is great for understanding what's most important in your data, but if you need all the details, Adkodas neural networks give you the most comprehensive analysis and predictive power using all your data.
Adkodas vs. the rest: rule extraction
Many practitioners of deep learning profess to offer to make machine learning your learning--but Adkodas offers true transparency into what the computer has learnt. This is what you need if you want to turn your data into science, or to prove that the AI controlling your equipment is making the right decisions, or just to understand what your data is saying.
This is your first step in trusting any AI and the results it gives.
Your data contains many patterns, or rules; finding the rules, the input, the relationships between input, or the values that matter most is what we do.
Our classification process was trialled at the Stanford University Physics Department in a PhD research project; Adkodas tech was able to automate the classification of different types of galaxies in large surveys. During the trial we also did rule extraction, so we had not only had an automated classification neural network, but were also able to extract information from the network of how to differentiate between galaxy types -- information that could be independently verified without going back to the neural network itself.
To understand Adkodas rule extraction and data analysis results, take a look for yourself: while there is no direct apples-to-apples comparison with Adkodas's unique technology, we do stack up our results with those from existing data sets.
Want to try our free service? We'll start by identifying and extracting for you a few key fields -- unlike other neural-network services, this can allow you to independently verify what those fields are doing to classify the data. All we'd like is a reference in your peer reviewed papers; if you'd like to know how much we can do for you, for free, or as a more in-depth analysis project, contact us!
Adkodas vs. the rest: talking technical about neural networks
Finding the rules in your data is one thing; being able to predict the output of new data is another. Rule extraction is about systematically removing data to find existing patterns. But if you are trying to apply new, previously unknown data to existing sets, our full neural networks are for you.
If you are familiar with neural networks (and specifically supervised learning), you'll want to compare our results to data sets trained with existing technologies. Our neural networks achieve over 90% accuracy rates on benchmark data sets from the UCI Machine Learning Repository (Abalone, German Credit, etc.), outperforms existing technology in predictive value, particularly for when large emergency responses are needed (Forest Fires), and attacks even one of the most difficult benchmarks (Two Spiral) with accuracy rates comparable to or surpassing existing technology while using vastly less time and resources. By comparison, the same data sets trained with standard backpropagation techniques typically get ~60% correct results, i.e. only somewhat better than chance.
While our automated results alone yield high accuracy rates compared to other feedforward neural network training algorithms, we draw on a variety of AI techniques to train our networks. We've solved many of the problems that assail ordinary neural networks, including
No local minima problem -- our networks are interested in relationships between data, not minimizing parameters; this method means cutting out the problem of local minima altogether.
Our networks use a deterministic training algorithm, meaning your data set is guaranteed to finish learning (we don't need an unknown, open-ended number of iterations to learn a data set).
No over-learning problem: redundant data that causes over-learning essentially provides no additional data. We replicate the biological ability to avoid over-learning, e.g. humans can be exposed to the alphabet as often as we like and still know how to use it, and we don't forget some letters just because we see them less often.
When we run the network, if the network output doesn't know the answer, it can tell you.Typical neural network outputs are bad at indicating between error and uncertainty. Humans either know the wrong answer, have learned the right answer or don't know.
The network will reproduce what it has learned with 100% accuracy.
While we certainly don't yet have all the answers to how learning occurs for humans or animals, Adkodas using biologically based neurons and network thereof means also having the side benefit of helping us to learn how biological learning works, all while doing better business, academic, and general data learning.
In particular, perfect recall of learned data is no obstacle to generalization: our networks generalize better than other feedforward network strategies. In addition, since the network recalls what it learns with 100% accuracy, you can have confidence that the rules from Adkodas' rule-extraction service are of a very high quality (since they must apply to all learned data, and not just a statistical "best fit") -- even before you independently verify the results.
Although our networks have been designed first and foremost as classification tools, particularly in initial development for, e.g., academics with large data sets and the need to automate their classification, we can also handle spatially dependent data (i.e. looking for patterns in data that depends on the order of the input fields of data). Because our preprocessor is primarily designed for classification data sets rather than spatial ones, you will likely want to convert your data to binary format first -- see our technical introduction for a more detailed discussion between "classification" and "spatial" data sets.
How Adkodas analysis works:
Adkodas is excited to offer, for a limited time, a free no-frills rule extraction and neural-network service.
Reach us with some basic details about your data set (learn what technical details we will need in data format), what results or patterns you are trying to learn from it, and a primary contact address where we can contact you with questions about the data.
We will make an initial analysis and then establish a dialogue about what you are looking to learn from your data, with follow-up questions from our end about what data fields represent; then we will lay out how we can help you -- as a free service or as a paid service.
Adkodas' data analysis services are completely confidential (except and unless you explicitly grant us permission to promote or advertise your work, see below); if your data are proprietary we can sign standard NDAs for additional security.
Our free service: If we provide a free service toward your not-for-profit research and the results are expected to be published, all we ask is you cite Adkodas™ (adkodas.com) in any publications referring to the data and its analysis. Once you cite us, if you desire and explicitly communicate such to us, we can provide a link to your paper or thesis, e.g. through social media. Naturally, if you are impressed with our free service and you can afford to donate, we are happy to accept Paypal.
Our paid service: If analysis of your data is to be used for any for-profit purpose, we will negotiate payment with you on a case-by-case basis, depending on the scale of the service you require as well as the scale of the deployment of the results. Our analysis is flexible and so are we! We want to help you unlock the hidden information in your data and understand its utility -- not only providing you with our exact learning algorithms, but also moving beyond merely providing an opaque "magic-eight-ball" neural-network prediction.