Thomas and Gunter
We had a chance to interview our CEO, Dr. Gunter Fischer who talked about how Brainalyzed happened and what his plans are for the future of this amazing swarm AI platform.

An interview with our CEO, Dr. Gunter Fischer

What is Brainalyzed?  

Brainalyzed is the creator of Brainalyzed Insight, the world’s first artificial swarm intelligence platform, tailored for people without a background in machine learning and artificial intelligence or programming. It enables industry experts with data and data hypothesis to build, deploy, and maintain powerful AI swarms, without the need for coding.  

What is the key differences between Brainalyzed and H2O, DataRobot, Google AutoML 

There are a lot of differences, but if I should pick the three most important ones, I would say it’s the way we create models with the aid of automated evolutionary AI design, the flexibility of the data pre-processing which gives the user the chance to combine data from distributed sources, and finally the transparency of our pricing. 

Could you elaborate on these differences? 

The key difference is how we create AI models. With Brainalyzed Insight, you can mimic the work of data scientists whose role is to find out which inputs have a predictive edge and which architecture is best suited for the prediction problem. This is a manual and time-consuming process. It requires a lot of expertise, especially if you’re not dealing with standard cases like object recognition or natural language processing, where you can find good off-the-shelf models in literature. 

All of the mentioned competitors merely do re-training of these well-known machine learning models with user-defined input data. However, for many cases, you have to start from scratch. For this purpose, in Brainalyzed Insight, we combine a custom-made genetic algorithm with deep learning.  

So, we are actively designing every aspect of the model, starting from inputs, architecture, to all of their trainable parameters. By doing this, we can improve the prediction accuracy with each new generation by combining the best models, using crossing over and mutation.  

The whole process is heavily parallelized with multiple running serverless processes on the cloud. Moreover, opposed to other AI providers, we can treat multiple competing prediction objectives at the same time. For example, this can be risk versus return for investment models or cost versus performance for models within a technical application. For these multi-objective cases, you will end up with multiple Pareto optimal models that are combined as an AI swarm at the end.  

Regarding data pre-processing, most AI providers rely on data processing outside the platform. However, data processing is really a key steppingstone to a working a model. Therefore, Brainalyzed Insight gives the user transparency about the data quality at every transformation step, especially if data from multiple data sources is joined into a combined training set. And if new data is available, the processing is conducted in exactly the same way every time.  

Finally, unlike other AI platforms, Brainalyzed Insight does not have hidden costs. For example, DataRobot charges extra for basic features such as time series. Our pricing is straightforward and simple – $24k for the first three months of POC and $8k each month for the subscription after that.   

Why is Swarm AI better than traditional AI? 

This boils down to the notion of diversification. Every model in the Swarm is potentially different in the inputs and architectures that are used. Therefore, not all models are impacted by outliers or noise in certain input variables and the overall prediction performance will be more stable.  

On the other hand, we allow single models in the swarm to specialize in specific patterns in the data. You can think of this as an individual in a society, that is only able to perform a single task but at a very high skill level. Finally, it is possible to dynamically change the Swarm composition, exchanging swarm models with low prediction accuracy with non-active models with higher prediction accuracy.  

What type of machine learning algorithms does your product use? How is it better? 

Our learning backend only uses deep neural networks as the type of machine learning model. We think that this is the most versatile model available at the moment since you can model most complex cases as well as really basic ones.  

On the other hand, the Architecture selection for these kinds of models is time-consuming and tricky. Opposed to gradient boostingr machines, deep neural networks have a vast variety of network components, that can be plugged together in countless ways.   

Moreover, every network component has additional parameters that need to be set sensibly. But if done right and combined with the right inputs, it will result in a very powerful machine learning model. Will take that process off the shoulder from the user, so that they can focus on proving whether or not their hypothesis works. 

How long before we start seeing results? 

It depends on the size of the data and the number of potential features, that you want to use for your use case. These factors determine how long a single model training will run. The genetic algorithm will generally run for around 10 to 40 Generations, depending on your convergence criterion, and it will train about hundreds to a couple of thousand different inputs and architecture combinations.  

Depending on the chosen parallelization this will take from a couple of hours to a couple of days. If the training is finished and the training server is shut down, you can then go ahead and choose the models you would like to use within the platform or deploy via API. 

Do we need a data scientist to get started with the AI project? 

The entire AI platform is designed for domain experts with no prior machine learning, programming, or coding experience. You only need data, a data hypothesis, and average excel level computer skills to get started. With a short demo, you can quickly grasp the concepts and be able to use the technology to create new AI models. 

How does it impact the workload of data scientists? 

Besides the architecture and input search, within the platform, we also provide a powerful preprocessing pipeline, which can be used to ingest data from multiple sources and perform various preprocessing tasks. This usually takes up 80% to 90% of the data scientist’s time. Our platform can help data scientists to speed up data preprocessing and model creation.  

We often see data scientist teams that are underwater when it comes to the number of projects they have to deal with. With this platform, they can focus on handcrafted models where this is needed and deliver the rest via Brainalyzed Insight. 

How efficiently does the platform integrate with thirdparty platforms and APIs? 

In essence, this boils down to the options given to the user to connect data to the platform, as well as getting results out of it. For data ingestion, we support direct file upload via CSV and database connection to relational databases which are state of the art for 95% of companies.  

The specialty in this regard is that you can connect to multiple databases at the same time and perform the whole joining operations in the preprocessing pipeline afterward. This saves you time and money that is necessary to establish a central Data Lake solution.  

Needless to say, that there are never enough data integration options in a platform, therefore we are constantly expanding the capabilities of the platform by connecting third-party data integration providers.   

For further processing of platform outputs, we provide a CSV download as well as a powerful rest API to integrate into countless other applications. This is relevant for people that want to visualize the automatic predictions via a power BI or tableau dashboard.  

Do I need coding experience to use the platform? 

No, you don’t need coding experience to use the platform. In my opinion, one of the major roadblocks to AI adoption is the lack of data scientists and people that are proficient with modern ML and data processing languages.  

Even if companies have invested in a data science team, we often see department silos with a lack of understanding and mistrust between data scientists and older members of staff with significant domain knowledge. This is understandable since they fear to be replaced at least to a certain extent by the resulting AI systems. To put them back in the driver seat even in the data-driven world, we have decided to make Brainalyzed Insight a 100% codeless platform.  

How does it impact revenue and costs? 

On one hand, there is a component of the efficiency increase of the workforce. That means you can accomplish more with the same staff headcount. If you’re working with data, you can only analyze a certain amount in a certain time with classical manual analytical approaches. But since the speed of data creation increases exponentially over time, it is only possible to analyze an ever-smaller portion of it. You simply cannot scale the workforce at a speed in which data is made available.  

With Brainalyzed Insight there’s technically no limit on the amount of data you can process and use to improve your business processes or create better products and services. This leads me to revenue, which is simply the other side of the same coin. If companies can provide better products and services, they will be superior to their competitors, have more satisfied customers, and in the end get more market share.  

What were you doing before you founded Brainalyzed? How would you have used Brainalyzed Insight to solve a problem in your previous organization? 

Before I started Brainalyzed I worked for more than 10 years in the renewable energy sector developing self-learning controllers for wind turbines. The goal was to use the time-series data of the wind turbine sensors to make the turbine adapt to local site conditions and improve the energy production within a given load envelope.   

We had the choice of over 300 technical sensors within the wind turbine and didn’t know how to combine them to solve the task. At that time artificial intelligence leapfrogged from being rather academic exercise to offering real business value with the rise of object detection in pictures.   

For these tasks, you could find a lot of good models in the literature already. However, if you have a new problem, that no one ever tackled before, you need to start with a clean slate. It is pretty difficult to search through all the available inputs and architecture options.  

Since the time I got a Ph.D. in technical optimization, I found it intriguing to just pose that problem to an optimizer. This is when the idea of automated evolutionary AI design was born and Brainalyzed happened. 

How do you see models optimizing over time? 

As mentioned, within the current system we are using an evolutionary algorithm to mimic the work of data scientists. But in the long run, I have the vision to use the training data of the hundreds of thousands of training we conduct to improve the process with an AI model itself.  

It will use the statistical properties of each input and the prediction targets, to suggest a good initial population and parameterization scheme, which is then used to further optimize the models with the current system.  

In other words, we want to improve the static nature of parameterization with data to drastically reduce the training time with the same or better prediction accuracy.  

Which cloud providers do you work with?  

From the start, Brainalyzed Insight was designed as a cloud-agnostic system. This is important first of all, to not be vendor locked in and to offer the customer the option to work with their preferred cloud provider. We have run the training system on Azure, AWS, and IBM already and I currently using the Hetzner cloud, which is the German provider with service centers in Germany and Finland.  

What types of data can be used for the models? 

Currently, Brainalyzed Insight can support numerical, text, and categorical inputs. Processing images is on the platform roadmap since the current learning system can deal with it.  

When it comes to the data structure, we support the processing of base data as well as time-series data, while most of the use cases we deal with the latter.   

One of the most powerful features of Brainalyzed Insight is its capability to deal with multiple stacked timeseries within the training, in case you have multiple entities, where each one produces timeseries data.   

What is your delivery model?  

We have made the experience that the best way forward, is to guide the customer through the initial steps with the learning system. Everybody knows nowadays what cloud-based storage or spreadsheet analysis is. Things look different for AI technology. It’s a great tool as soon as you know how to use it 

We often face the challenge, that we need to educate customers on how to best derive value from such systems. But as soon as they see the numerous possibilities, they are amazed. Knowing all that, we usually engage with an initial proof of concept to look at the available data and the business challenge that needs to be solved, create the initial models together and then hand over the models to the customer with a subscription of the platform. 

Is your platform scalable?  

Yes indeed! And it needs to be if you want to serve an ever-growing customer base with the ever-increasing complexity of use cases.  Luckily with today’s Cloud technology, scaling becomes rather a commodity than an asset.   

What’s a recent (exciting) project you’ve worked on/are working on? 

In essence, every project is exciting, since we do not know the result in advance, and if we even get data that contains the patterns to solve the case. Since I have an engineering background, I’m excited by models that can approximate difficult scientific rules from physics. One challenging problem in that respect was the design of self-learning controllers.  

On the other hand, it is also fascinating how one can model human behavior with such a system. The question of emotions in the financial markets or the probability of people to churn by just looking at their data from web tracking. We’ve worked in all these areas, which is the advantage of the dataagnostic nature of the learning system.  

Tell me something about security and data privacy policies? 

This is an important point for us as well as for customers. To me, it’s a hygiene factor for digital businesses, considering the rising crime rates on the web. We are putting a lot of effort into making our systems safe, by adopting the latest encryption standards using virtual private networks and implementing a rigorous access control level system.   

To make things even safer, we are working on a separate encryption layer, that is activated even before the data is sent from the customer to our backend. So, in case we would have a man in the middle attack the data that is captured is of no value to the other party. This is only possible since we are not interested in the data itself but only in the statistical properties.  

Who owns the AI models?  

Since we usually do not provide the data for the models, they are the sole property of the customer and we are not reselling them in any form. The only data that we keep is data about the input statistics, the learning progress, and the optimization history, to do the job better next time. 

Subscribe to our newsletter

AI for anomaly

Want to know more about AI but don’t know where to start?

Get a free copy of our whitepaper.

Related articles

AI

Responsible AI

Responsible AI is a framework that registers how an organization is addressing the challenges around artificial intelligence (AI) from an ethical and legal perspective.

Read More »

AI platform for the world’s
data-driven companies