About the Customer
Revenue (20B$ +)
We took the challenge of predicting optimal parameters for a device used in medical research.
- The device has many parameters which are determined experimentally.
- A lot of manual effort has to be involved to set the proper parameters to get results.
- In some cases, the samples they use might be rare or it takes much time for sample preparation, and cannot afford to test so many times with different sets of parameters.
- So, there should be some automated process to find the correct parameters, which Machine learning or Deep Learning can solve.
Our Starting Point
- The parameter values are tried and tested manually.
- Testing will continue until the optimal parameters are found.
- In the case of rare samples, the sample will get wasted when tested with so many values.
Machine Learning Potential
- The model developed will increase the range of research and experimentation, while reducing the time and effort taken in sample preparation and other steps for scientists.
- Increased selectivity of parametres results in low cost and improved ease of operation.
- The sample used could be rare, valuable and costly, as the sample is rare there can only limited experiments that can be done.
- Time and effort and cost spent on each experiment is high.
- Data-driven and AI-driven approach to solving the problem
- Built a deep learning model to predict ideal parameters.
- Use predera’s AIQ platform to easily build, iterate, deploy and monitor the Deep learning model
- Schedule to train the model at regular intervals to improve accuracy
The required work is feature extraction and deep learning models to optimize the parameters.
Fig: Model Architecture
- This is a deep learning problem as there it non-linearity, which deep learning can do well.
- Neural networks can identify hidden patterns in the data without exclusive feature engineering.
- The traditional algorithms cannot perform well or separate out the signals when there is non-linearity, which “Artificial Neural Networks” also called deep learning can do.
- Traditional algorithms need most of the features identified by an expert and then hand-coded as per domain and data type.
- The performance of most machine learning algorithms depends on how accurately the features are identified and extracted.
- Deep learning algorithms try to learn high-level features from data which is a major step ahead of machine learning.
- Deep learning reduces the task of developing new feature extractor for every problem.
An artificial neural network is developed to predict the ideal parameters.
The network takes protein data as input and the input is passed through hidden layers of the neural network and finally to output layer which gives the predicted optimal parameters.
The Predera team experimented different algorithms with different representation of input features and developed a model to predict optimal parameters which reduce the window size of the parameter values to be experimented.
The client was able to easily run predictions and explore more.
The model achieved 95% accuracy(no.of correct predictions) with error rate of 20 volts, 75% for error rate 10 volts and 50% for error rate 5 volts.