Figure 3

The signal acquisition was performed using a time-domain Finite State Automaton (FSA). This was followed by adaptive pre-filtering using a wavelet-domain FSA. Feature extraction on those acquired channel blockades was done by Hidden Markov Model (HMM) processing; and classification was done by Support Vector Machine (SVM). The optimal SVM architecture is shown for classification of five DNA hairpin molecules labeled 9CG, 9GC, 9TA, 9AT, and 8GC (the number denotes the stem length in base-pairs and the two-base entry denotes the 5'-3' termini). The linear tree multi-class SVM architecture benefits from strong signal skimming and weak signal rejection along the line of decision nodes. Scalability to larger multi-class problems is possible since the main on-line computational cost is at the HMM feature extraction stage. The accuracy shown is for single-species mixture identification upon completing the 15^th single molecule sampling/classification (in approx. 6 seconds).